New generation of Liquid Foundation Models featuring a tri-hybrid architecture and multimodal encoders. The family includes six ~2B-parameter models covering text (256k context), vision-language with video, audio (18 languages, tool calls), omni, and two decision models with typed outputs. Open weights, fine-tunable on a single GPU, optimized for 14 processor classes and on-device deployment.
LFM3 is Liquid AI’s family of open-weight foundation models, built around a shared LFM3-1.7B-Base backbone. The family includes text, vision-language, audio, omni, and decision models, with variants specialized for different inputs and outputs. Liquid AI describes the models as designed for local deployment and fine-tuning, including on a single GPU; the text model has a documented 256K-token context window. The models are available through model repositories, Liquid AI’s API, and its Playground. Documented uses include tool calling, speech interaction, video understanding, and structured decisions.
